Pith. sign in

REVIEW 4 major objections 5 minor 37 references

Diffusion Models for Inverse Problems in the Exponential Family

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that, for one-parameter exponential-family observations, the intractable likelihood score of a diffusion solver becomes tractable through a conjugacy-backed closed-form evidence approximation, enabling Poisson and…

desk verdict Clever and genuinely new evidence trick, but the independent-conjugate approximation is unvalidated at image scale; send to peer review with a request for a targeted validation experiment. read the letter →

arxiv 2502.05994 v1 pith:7MFHVABU submitted 2025-02-09 stat.ML cs.LG

Alessandro Micheli , Mélodie Monod , Samir Bhatt This is my paper · ORCID
classification stat.MLcs.LG MSC 62F1560J6062M3068T07
keywords diffusionmodelsinverseproblemsexponentialfamilylikelihoodscoreevidencetrickconjugateprioramortizedvariationalinferencescore-basedCoxprocess
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces the evidence trick, a way to compute the likelihood score term that blocks diffusion-model solvers from handling non-Gaussian observations. The claim is that for any one-parameter exponential-family likelihood, the intractable integral over the reverse diffusion process can be approximated in closed form by placing a natural conjugate prior on the likelihood parameter and differentiating the resulting evidence. If correct, this removes the Gaussian-noise assumption that has confined diffusion-based inverse solvers, letting an ImageNet-trained diffusion prior act directly as a prior over Poisson intensities or disease-prevalence maps. The authors demonstrate the method on synthetic benchmarks against MCMC, on recovering image-like intensities from point counts, and on malaria prevalence estimation, where it matches the accuracy of established spatial models.

What carries the argument

The evidence trick: for a one-parameter exponential-family likelihood $p_{\mathbf{y}|\boldsymbol{\theta}}$, the natural conjugate prior $q_{\boldsymbol{\theta}|\boldsymbol{\zeta}}(\boldsymbol{\theta}|\boldsymbol{\zeta})$ is an exponential-family distribution with sufficient statistics $(\eta(\boldsymbol{\theta}), -A_{\mathbf{y}}(\eta(\boldsymbol{\theta})))$, and the integral $\int p_{\mathbf{y}|\boldsymbol{\theta}} q_{\boldsymbol{\theta}|\boldsymbol{\zeta}} \, d\boldsymbol{\theta}$ has the closed-form ratio $h_{\mathbf{y}}(\mathbf{y}) \exp(-A_{\boldsymbol{\theta}}(\nu, \tau)) / \exp(-A_{\boldsymbol{\theta}}(T_{\mathbf{y}}(\mathbf{y})+\nu, \tau + N \mathbf{1}_d))$. Differentiating the logarithm of this ratio gives the approximate likelihood score, so this ratio carries the entire argument. The hyperparameters $\boldsymbol{\zeta}(\mathbf{x}_t)=(\nu(\mathbf{x}_t),\tau(\mathbf{x}_t))$ are predicted by a neural-network inference network trained by minimizing an amortized variational objective that, by Theorem 3.5, equals $A_{\boldsymbol{\theta}}(\boldsymbol{\zeta}_\rho(\mathbf{x}_t,t)) - \boldsymbol{\zeta}_\rho(\mathbf{x}_t,t)^\top T_{\boldsymbol{\theta}}(g^{-1}(\mathbf{x}_0))$ under expectations over priors and forward noise, a step that avoids the reverse-process expectation entirely. A deterministic link function $g$ completes the machinery by mapping the latent $\mathbf{x}_0$ onto the likelihood parameter domain, for example the exponential map for Poisson rates and the sigmoid for probabilities.

What would settle it

On a low-dimensional hierarchical model with an exactly computable or MCMC-accessible posterior, compare the evidence-trick score against a high-accuracy Monte Carlo estimate of $\nabla_{\mathbf{x}_t} \log p_{\mathbf{y}|\mathbf{x}_t}$ at fixed noise levels $t$; a disagreement that grows as the likelihood becomes more informative (e.g., Poisson rates approaching zero) would indicate the variational approximation breaks where it is needed.

Watch

Extended reading notes

Core claim

In the author's formulation, the score of the posterior at diffusion time $t$ splits into the prior score plus $\nabla_{\mathbf{x}_t} \log p_{\mathbf{y}|\mathbf{x}_t}(\mathbf{y}|\mathbf{x}_t)$, and the paper's central claim is that this second term is accurately approximated by replacing the intractable posterior $p_{\boldsymbol{\theta}|\mathbf{x}_t}$ with a conjugate exponential-family variational distribution $q_{\boldsymbol{\theta}|\boldsymbol{\zeta}(\mathbf{x}_t)}$. Conjugacy then collapses the evidence integral to a closed-form ratio of log-partition functions, Equation (10), and the hyperparameters $\boldsymbol{\zeta}(\mathbf{x}_t)$ are learned by an amortized variational objective (Theorem 3.5) that is trained on prior samples alone, without Tweedie's formula. The gradient of this closed-form log-evidence with respect to $\mathbf{x}_t$ is the approximate likelihood score that drives the reverse SDE, yielding posterior samples of the latent variable $\mathbf{x}_0$ and, through the inverse link function, of the likelihood parameter $\boldsymbol{\theta}$.

Load-bearing premise

The conjugate exponential-family distribution learned by the inference network must approximate the true reverse posterior closely enough at every diffusion time for the closed-form evidence gradient to behave like the true likelihood score, even though the network is trained on prior samples and then evaluated on posterior-conditioned noisy states.

Editorial extensions

If this is right

  • Inverse problems with count data — epidemic case numbers, astronomical photon counts, medical imaging — become solvable by training a score network on a prior and using the same likelihood-agnostic sampler for Poisson observations.
  • The variational objective is trained purely on prior samples and does not require the observations, so a single trained inference network applies to any missing-data pattern and any dataset for a given likelihood family.
  • A diffusion prior learned on natural images can act as a spatial prior for Poisson process intensities, turning the method into a 'score-based Cox process' that interpolates sparse point counts into full intensity maps.
  • Posterior uncertainty (credible intervals) matches MCMC in the synthetic benchmark, in contrast to DPS, which under-covers uncertainty at matched point estimates.
  • For disease mapping, the diffusion-prior approach achieves mean absolute error comparable to a GMRF-based state-of-the-art on the malaria PfPR data, suggesting a scalable alternative to spatial statistics when full-rank Gaussian inference is infeasible.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same conjugacy recipe should apply to every family in the paper's table (Gamma, Weibull, Pareto, Negative Binomial, Geometric), offering a plug-and-play likelihood score where Gaussian surrogates do not exist; this is a direct corollary of the one-parameter exponential-family result.
  • A natural testable extension is to increase the flexibility of the variational family (e.g., mixture or normalizing-flow conjugate families) and measure whether the score approximation improves at high noise levels, where the independent conjugate family is likely least faithful.
  • The link-function perspective connects to generalized linear models; one could use the evidence trick to build diffusion priors for GLM coefficients, bringing score-based sampling to regression settings with non-Gaussian responses.
  • Numerical stability (gradient clipping, SNR tuning) noted in the paper suggests the evidence score may be ill-conditioned for very small counts; a practical follow-up is to reparameterize the log-partition functions or use stabilized log-gamma and log-beta computations.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes the 'evidence trick,' a closed-form approximation to the likelihood score for diffusion-model-based posterior sampling when observations come from a one-parameter exponential family. The key idea is to approximate the intractable reverse conditional p(θ|x_t) by a conjugate exponential-family variational distribution q(θ|ζ(x_t)), whose hyperparameters are learned by an amortized inference network trained only on prior samples. This makes the evidence integral in Eq. (9) analytically tractable, yielding Eq. (10), and the likelihood score is then obtained by automatic differentiation. The method is demonstrated on a 1D Gaussian-process benchmark with MCMC comparisons, on image-scale Cox process intensity recovery with an ImageNet prior, and on malaria prevalence mapping in Sub-Saharan Africa with Binomial observations, where it is compared to a GMRF baseline.

Significance. If the central approximation is valid, the evidence trick is a genuinely useful contribution: it extends score-based diffusion solvers for inverse problems beyond Gaussian likelihoods to Poisson, Binomial, and other exponential-family observations, without requiring observations to be encoded into the inference network. The mathematical derivation in Section 3 and Appendix I is internally consistent under the stated assumptions, and the design of training the inference network solely on prior samples is clean and avoids fitting to the target posterior. The paper also ships code and reproduces a real-world disease-mapping experiment, which is a strength. However, the load-bearing approximation—that a product-form conjugate q approximates p(θ|x_t) well enough for image-scale correlated priors—is not validated by any experiment that compares Eq. (10) against the integral it approximates, and the empirical evaluation is largely qualitative. The significance is therefore conditional: the method is promising and likely correct as a derivation, but the evidence presented does not yet establish that it works as claimed at scale.

major comments (4)
  1. [Sec. 3.3, Eq. (9)-(10)] The central approximation is that p(θ|x_t) can be replaced by an independent conjugate-family distribution q(θ|ζ(x_t)) = ∏_j q_j(θ_j|ζ_j(x_t)). The KL objective in Lemma 3.4, and hence the amortized training in Theorem 3.5 and Eq. (13), only matches the marginal expected sufficient statistics E[η(θ_j)|x_t] and E[A_y(η(θ_j))|x_t]; it does not match joint structure. For a correlated image prior, p(θ|x_t) is not factorized at finite noise levels, because the forward noising transition p(x_t|x_0) does not eliminate spatial dependence. The evidence integral in Eq. (9) is a joint expectation of a product over coordinates, and the independent-q approximation replaces it with a product of one-dimensional expectations, which is exact only when p(θ|x_t) factorizes. This is a misspecification error that would persist even with a perfectly trained amortization network, and no experiment isolates it. I ask the authors to add a controlled experiment—e.g., a moderate-dimensional correlated prior where the true integral in Eq. (9) can be evaluated by MCMC or sequential Monte Carlo—and report the error of the Eq. (10) approximation as a function of noise level t and correlation strength.
  2. [Sec. 4.1 and Appendix G.1] The one-dimensional benchmark is presented as demonstrating that the posterior 'closely aligns' with MCMC, but the results are only qualitative: Figures A5-A9 show posterior medians and 95% credible intervals, with no numerical metrics such as coverage of the true value, average interval score, RMSE of posterior medians, or marginal KL divergence to the MCMC posterior. The claim that DPS 'fails to accurately capture both the credible intervals and the point estimates' is likewise supported only by visual inspection. Quantitative comparison is needed to substantiate the central claim that the method produces calibrated posterior approximations, especially because the Appendix reports that for the high-rate Poisson case the proposed method is not visually distinguishable from MCMC but no error measure is given.
  3. [Sec. 4.2-4.3 and Appendix G.2-G.3] For the image-scale Cox process experiments, there is no ground-truth posterior and no held-out predictive criterion: the results are posterior medians and interquartile ranges only. Since the method is solving an inverse problem, at minimum a held-out log-likelihood or predictive calibration of the Poisson counts should be reported. For the malaria experiment, the comparison to GMRF rests on a single MAE value (0.1207 vs 0.1225, Appendix G.3), with no statement of whether this is computed on the training or test set and no confidence interval or paired comparison; the posterior predictive coverage reported in Appendix G.3 (78% test-set coverage for the proposed method vs 43% for GMRF) is a more informative statistic and should be brought into the main text. As it stands, the manuscript's headline claims about 'effectively performing Bayesian inference' and 'performing competitively' are not supported by sufficient quantitative evaluation.
  4. [Sec. 3.4, Eq. (13)] The inference network ζ_ρ is trained on pairs (x_t, x_0) drawn from the prior forward process, but at sampling time it is evaluated on x_t trajectories generated by the posterior reverse process. Because the posterior-conditioned distribution of x_t differs from the prior forward noising distribution, the amortized network operates under a distribution shift that is not analyzed. This is distinct from the factorization issue in the first comment and would remain even if the variational family were expressive enough to capture p(θ|x_t). I ask for at least one diagnostic: in the 1D GP setting, where the MCMC conditional is available, evaluate the learned q(θ|ζ(x_t)) at posterior-sampled x_t and compare its moments or density against the MCMC conditional, and report the difference. This would directly test whether the amortized network generalizes to the posterior path.
minor comments (5)
  1. [Appendix E, sampling algorithm] The expression for g_likelihood writes '∇_x_t (h_y(y) − A_θ(...) + A_θ(...))'; since Eq. (10) is a log-density, the first term should be log h_y(y). For Poisson and Binomial likelihoods the gradient of log h_y(y) is zero, so the final gradient is unaffected, but the notation is misleading.
  2. [Appendix C and Table A1] The Gamma and Beta conjugate priors require constraints on the hyperparameters (e.g., shape parameters positive) for the log-partition functions to be finite. The implementation section does not explain how the neural network output ζ_ρ(x_t, t) is constrained to satisfy these conditions; please clarify.
  3. [Sec. 4.2] The text says that 80% of grid cells were allocated to the training set and 20% to the test set, but the method does not train on observations. Please state clearly what is evaluated on the test set and what metric is used.
  4. [Appendix E, architecture] The phrase '100,000 epochs' is likely intended to mean '100,000 iterations' for a batch size of 128; please correct the terminology, since an epoch would imply multiple passes over the entire ImageNet training set.
  5. [Sec. 4.3] The main text reports only medians and quantile maps for the malaria experiment; the MAE comparison and posterior predictive coverage are relegated to Appendix G.3. At least the test-set coverage comparison should be moved to the main text, since it is the strongest quantitative evidence for calibration.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the evidence trick is a variational approximation trained on prior samples and benchmarked against external MCMC and GMRF baselines.

full rationale

The paper's derivation is self-contained and does not reduce to its own inputs. The evidence trick (Eqs. 9-10) is a stated approximation: it replaces the intractable reverse posterior p_theta|x_t with an independent natural-conjugate exponential-family q_theta|zeta(x_t), then evaluates the likelihood evidence in closed form via the standard conjugacy identity in Proposition A.8. The variational hyperparameters zeta_rho(x_t,t) are trained exclusively on prior samples (x0 ~ p_x0, xt ~ p_xt|x0) using the amortized objective in Eq. (13), which is obtained from Eq. (12) by the exact Bayes identity p(xt)p(x0|xt) = p(xt|x0)p(x0), as proven in Theorem 3.5 with the integrability check in Lemma I.1. The likelihood score then enters through the conjugate update (nu + Ty(y), tau + N 1_d); this is a standard Bayesian update, not a reuse of the target posterior or of the reported experimental comparisons. Because the observations y are never used to train the inference network, there is no fitted-input-called-prediction pattern, and no step defines the claimed output in terms of itself. The paper's main weakness, that an independent product-form q may not capture the correlated reverse posterior for image priors, is a genuine approximation or misspecification concern, not a circularity: the target integral in Eq. (5) is not constructed from the approximating family, and the approximation error is not hidden by a fitted parameter. The 1D experiment is validated against MCMC ground truth, and the malaria experiment is benchmarked against an external GMRF/INLA baseline with posterior predictive checks, so the central claim has independent empirical content. The only author-overlap references (Bhatt et al. for PfPR data and disease-mapping context) are external data and benchmark citations and are not load-bearing for the derivation.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The method explicitly assumes factorized one-parameter exponential family likelihoods (Assumptions 3.1-3.2) and a smooth bijective link (Assumption 3.3). The key extra assumption is that a conjugate exponential-family variational distribution can approximate the reverse posterior p_theta|xt well enough for the evidence integral to provide a useful likelihood score; this is not theoretically bounded or quantitatively validated. Free parameters include the link scaling s=5 and sampling hyperparameters chosen by hand.

free parameters (3)
  • s (link scaling for malaria binomial model) = 5
    Chosen by hand in Section F.3 to stretch the sigmoid link so the ImageNet prior covers Binomial probabilities from about 0.001 to 0.99; not fitted to malaria data but affects the posterior scale.
  • signal-to-noise ratio r in predictor-corrector sampler = 0.1
    Selected by experimentation in Appendix E for stable posterior sampling; affects all experimental results.
  • gradient clipping threshold = [-10, 10]
    Hand-set to prevent numerical instabilities in the likelihood score gradient, as described in Appendix E.
assumptions (6)
  • domain assumption Assumption 3.1: Components of y are conditionally independent given theta, and each y_i,j depends only on theta_j.
    Defines the scope to factorized exponential family likelihoods; used throughout Section 3.3.
  • domain assumption Assumption 3.2: The likelihood belongs to the univariate one-parameter exponential family.
    Restricts the method to one-parameter families such as Poisson and Binomial.
  • standard math Assumption 3.3: The link function g is continuously differentiable, one-to-one, and has nonzero derivative.
    Needed for the change of variables between theta and x0; standard in GLM theory.
  • domain assumption The reverse diffusion marginal at time 0 equals the data prior px0 and the score network is accurate.
    Used in Theorem 3.5 to replace the reverse conditional by forward conditionals p_xt|x0 px0; standard but approximate with finite-capacity score networks.
  • ad hoc to paper The conjugate variational family q_theta|zeta(xt) approximates p_theta|xt sufficiently well for the evidence integral to give a useful likelihood score.
    Core approximation of the paper; no bound or quantitative validation is provided for image-scale priors.
  • standard math Moment conditions and Lipschitz continuity of zeta in Theorem 3.5 hold.
    Technical conditions for the Fubini exchange in the proof; assumed or satisfied for bounded image priors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Diffusion Models for Inverse Problems in the Exponential Family." pith.science (2026). https://pith.science/paper/7MFHVABU

@misc{pith2026250205994,
  author       = {Pith},
  title        = {Pith review of: Diffusion Models for Inverse Problems in the Exponential Family},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7MFHVABU}},
  note         = {Machine review of arXiv:2502.05994}
}
read the original abstract

Diffusion models have emerged as powerful tools for solving inverse problems, yet prior work has primarily focused on observations with Gaussian measurement noise, restricting their use in real-world scenarios. This limitation persists due to the intractability of the likelihood score, which until now has only been approximated in the simpler case of Gaussian likelihoods. In this work, we extend diffusion models to handle inverse problems where the observations follow a distribution from the exponential family, such as a Poisson or a Binomial distribution. By leveraging the conjugacy properties of exponential family distributions, we introduce the evidence trick, a method that provides a tractable approximation to the likelihood score. In our experiments, we demonstrate that our methodology effectively performs Bayesian inference on spatially inhomogeneous Poisson processes with intensities as intricate as ImageNet images. Furthermore, we demonstrate the real-world impact of our methodology by showing that it performs competitively with the current state-of-the-art in predicting malaria prevalence estimates in Sub-Saharan Africa.

Figures

Figures reproduced from arXiv: 2502.05994 by the authors.

Figure 1
Figure 1. Illustration of the approach using Diffusion Models for Inverse Problems in the Exponential Family. By leveraging the posterior score ∇xt pxt|y(xt|y), a reverse stochastic differential equation (SDE) can be solved to generate posterior samples of the latent variable x0 from noise. Posterior samples of the parameter θ are obtained by applying a deterministic inverse link function. The prior score function, ∇xt pxt (x… view at source ↗
Figure 2
Figure 2. Hierarchical Probabilistic Model. The dotted arrow represents a deterministic relationship, while the solid arrow indi￾cates a probabilistic relationship. 3.3. The Evidence Trick To approximate the likelihood py|xt (y|xt) as defined in (5), we propose a simple yet effective approach that we call the evidence trick. Given the assumption that the likelihood py|θ(y|θ) belongs to the exponential family, there always exi… view at source ↗
Figure 3
Figure 3. Score-Based Cox Process Results. (a) (Left) True Cox Process intensity from the ImageNet validation set, transformed using an exponential link function. (Right) Median of the estimated Cox Process intensity posterior distribution using the Score-Based Cox Process method. (b) (Left) True Cox Process Intensity from Sentinel-2 Satellite Imagery of Manhattan, New York City (Right) Median of the estimated Cox Process int… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Prevalence of Malaria in Sub-Saharan Africa Results. (a) Empirical PfPR. (b) Median of the estimated PfPR posterior distribution. (c) 25% quantile of the estimated PfPR posterior distribution. (d) 75% quantile of the estimated PfPR posterior distribution. The inset plo…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 30 canonical work pages

  1. [1]

    P., Murray, I., and MacKay, D

    Adams, R. P., Murray, I., and MacKay, D. J. C. Tractable N onparametric B ayesian I nference in P oisson P rocesses with G aussian P rocess I ntensities. In Proceedings of the 26th Annual International Conference on Machine Learning, ICML '09, pp.\ 9–16, New York, NY, USA, 2009. Association for Computing Machinery

    work page 2009

  2. [2]

    Anderson, B. D. Reverse-time diffusion equation models. Stochastic Processes and their Applications, 12 0 (3): 0 313--326, 1982

    1982

  3. [3]

    Conditional I mage G eneration with S core- B ased D iffusion M odels, 2021

    Batzolis, G., Stanczuk, J., Schönlieb, C.-B., and Etmann, C. Conditional I mage G eneration with S core- B ased D iffusion M odels, 2021. ar X iv preprint ar X iv:2111.13606

    arXiv 2021

  4. [4]

    J., Cameron, E., Bisanzio, D., Mappin, B., Dalrymple, U., Battle, K

    Bhatt, S., Weiss, D. J., Cameron, E., Bisanzio, D., Mappin, B., Dalrymple, U., Battle, K. E., Moyes, C. L., Henry, A., Eckhoff, P. A., Wenger, E. A., Briët, O., Penny, M. A., Smith, T. A., Bennett, A., Yukich, J., Eisele, T. P., Griffin, J. T., Fergus, C. A., Lynch, M., Lindgren, F., Cohen, J. M., Murray, C. L. J., Smith, D. L., Hay, S. I., Cibulskis, R. ...

    work page 2000

  5. [5]

    R., Weiss, D

    Bhatt, S., Cameron, E., Flaxman, S. R., Weiss, D. J., Smith, D. L., and Gething, P. W. Improved prediction accuracy for disease risk mapping using G aussian process stacked generalization. Journal of the Royal Society Interface, 14 0 (134), 2017

    work page 2017

  6. [6]

    Probability and Measure

    Billingsley, P. Probability and Measure. Wiley Series in Probability and Statistics. Wiley-Blackwell, 2012

    work page 2012

  7. [7]

    Pattern Recognition and Machine Learning

    Bishop, C. Pattern Recognition and Machine Learning. Information Science and Statistics. Springer, New York, NY, 2016

    work page 2016

  8. [8]

    Boys, B., Girolami, M., Pidstrigach, J., Reich, S., Mosca, A., and Akyildiz, O. D. Tweedie M oment P rojected D iffusions F or I nverse P roblems, 2024. ar X iv preprint ar X iv:2310.06721

    arXiv 2024

Show all 37 references
  1. [9]

    T., Klasky, M

    Chung, H., Kim, J., Mccann, M. T., Klasky, M. L., and Ye, J. C. Diffusion P osterior S ampling for G eneral N oisy I nverse P roblems. In The Eleventh International Conference on Learning Representations, 2023

    2023

  2. [10]

    Cox, D. R. Some S tatistical M ethods C onnected with S eries of E vents. Journal of the Royal Statistical Society: Series B (Methodological), 17 0 (2): 0 129--157, 1955

    1955

  3. [11]

    C., Milanfar, P., Dimakis, A

    Daras, G., Chung, H., Lai, C.-H., Mitsufuji, Y., Ye, J. C., Milanfar, P., Dimakis, A. G., and Delbracio, M. A S urvey on D iffusion M odels for I nverse P roblems, 2024. ar X iv preprint ar X iv:2410.00083

    2024 arXiv

  4. [12]

    Target S core M atching, 2024

    De Bortoli, V., Hutchinson, M., Wirnsberger, P., and Doucet, A. Target S core M atching, 2024. ar X iv preprint ar X iv:2402.08667

    2024 arXiv

  5. [13]

    Tweedie’s F ormula and S election B ias

    Efron, B. Tweedie’s F ormula and S election B ias. Journal of the American Statistical Association, 106 0 (496): 0 1602–1614, 2011

    2011

  6. [14]

    Compositional S core M odeling for S imulation- B ased I nference

    Geffner, T., Papamakarios, G., and Mnih, A. Compositional S core M odeling for S imulation- B ased I nference. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pp.\ 11098--11116. PMLR, 2023

    2023

  7. [15]

    London Building Stock Model (LBSM)

    Greater London Authority (GLA), Mayor of London . London Building Stock Model (LBSM) . Available at ://data.london.gov.uk/dataset/london-building-stock-model, 2025

    2025

  8. [16]

    J., Datta, A., Finley, A

    Heaton, M. J., Datta, A., Finley, A. O., Furrer, R., Guinness, J., Guhaniyogi, R., Gerber, F., Gramacy, R. B., Hammerling, D., Katzfuss, M., Lindgren, F., Nychka, D. W., Sun, F., and Zammit-Mangion, A. A C ase S tudy C ompetition A mong M ethods for A nalyzing L arge S patial ...

    2018

  9. [17]

    Denoising D iffusion P robabilistic M odels

    Ho, J., Jain, A., and Abbeel, P. Denoising D iffusion P robabilistic M odels. In Advances in Neural Information Processing Systems, volume 33, pp.\ 6840--6851, 2020

    2020

  10. [18]

    J., Duvenaud, D

    Johnson, M. J., Duvenaud, D. K., Wiltschko, A., Adams, R. P., and Datta, S. R. Composing graphical models with neural networks for structured representations and fast inference. In Advances in Neural Information Processing Systems, volume 29, 2016

    2016

  11. [19]

    and Simoncelli, E

    Kadkhodaie, Z. and Simoncelli, E. P. Stochastic solutions for linear inverse problems using the prior implicit in a denoiser. In Beygelzimer, A., Dauphin, Y., Liang, P., and Vaughan, J. W. (eds.), Advances in Neural Information Processing Systems, 2021

    2021

  12. [20]

    SNIPS : S olving N oisy I nverse P roblems S tochastically

    Kawar, B., Vaksman, G., and Elad, M. SNIPS : S olving N oisy I nverse P roblems S tochastically. In Advances in Neural Information Processing Systems, 2021

    2021

  13. [21]

    D enoising D iffusion R estoration M odels

    Kawar, B., Elad, M., Ermon, S., and Song, J. D enoising D iffusion R estoration M odels. In Advances in Neural Information Processing Systems, 2022

    2022

  14. [22]

    An explicit link between G aussian fields and G aussian M arkov random fields: the stochastic partial differential equation approach

    Lindgren, F., Rue, H., and Lindström, J. An explicit link between G aussian fields and G aussian M arkov random fields: the stochastic partial differential equation approach. Journal of the Royal Statistical Society Series B: Statistical Methodology, 73 0 (4): 0 423--498, 2011

    2011

  15. [23]

    A., Lucas, T

    Pfeffer, D. A., Lucas, T. C. D., May, D., Harris, J., Rozier, J., Twohig, K. A., Dalrymple, U., Guerra, C. A., Moyes, C. L., Thorn, M., Nguyen, M., Bhatt, S., Cameron, E., Weiss, D. J., Howes, R. E., Battle, K. E., Gibson, H. S., and Gething, P. W. malariaAtlas : an R interfac...

    2018

  16. [24]

    Rasmussen, C. E. and Williams, C. K. I. Gaussian Processes for Machine Learning. The MIT Press, 2005. ISBN 026218253X

    2005

  17. [25]

    and Fitzpatrick, P

    Royden, H. and Fitzpatrick, P. Real Analysis. Prentice Hall, 4th edition, 2010. ISBN 978-0-13-143747-0

    2010

  18. [26]

    Learning D iffusion P riors from O bservations by E xpectation M aximization

    Rozet, F., Andry, G., Lanusse, F., and Louppe, G. Learning D iffusion P riors from O bservations by E xpectation M aximization. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  19. [27]

    and Held, L

    Rue, H. and Held, L. Gaussian Markov Random Fields: Theory and Applications, volume 104 of Monographs on Statistics and Applied Probability. Chapman and Hall/CRC, 2005

    2005

  20. [28]

    Approximate B ayesian inference for latent G aussian models by using integrated nested L aplace approximations

    Rue, H., Martino, S., and Chopin, N. Approximate B ayesian inference for latent G aussian models by using integrated nested L aplace approximations. Journal of the Royal Statistical Society Series B: Statistical Methodology, 71 0 (2): 0 319--392, 2009

    2009

  21. [29]

    Neural S core E stimation: L ikelihood- F ree I nference with C onditional S core B ased D iffusion models

    Simons, J., Sharrock, L., Liu, S., and Beaumont, M. Neural S core E stimation: L ikelihood- F ree I nference with C onditional S core B ased D iffusion models. In Fifth Symposium on Advances in Approximate Bayesian Inference, 2023

    2023

  22. [30]

    Pseudoinverse- G uided D iffusion M odels for I nverse P roblems

    Song, J., Vahdat, A., Mardani, M., and Kautz, J. Pseudoinverse- G uided D iffusion M odels for I nverse P roblems. In International Conference on Learning Representations, 2023

    2023

  23. [31]

    P., Kumar, A., Ermon, S., and Poole, B

    Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. Score- B ased G enerative M odeling through S tochastic D ifferential E quations. In International Conference on Learning Representations, 2021

    2021

  24. [32]

    P y S tan, a P ython interface to S tan, a package for B ayesian inference

    Stan Development Team . P y S tan, a P ython interface to S tan, a package for B ayesian inference. Available at https://pystan.readthedocs.io/en/latest/, 2025

    2025

  25. [33]

    Stein, M. L. Limitations on low rank approximations for covariance matrices of spatial data. Spatial Statistics, 8: 0 1--19, 2014

    2014

  26. [34]

    N., Kaiser, L

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L. u., and Polosukhin, I. Attention I s A ll Y ou N eed. In Advances in Neural Information Processing Systems, volume 30, 2017

    2017

  27. [35]

    A C onnection B etween S core M atching and D enoising A utoencoders

    Vincent, P. A C onnection B etween S core M atching and D enoising A utoencoders. Neural Computation, 23 0 (7): 0 1661--1674, 2011

    2011

  28. [36]

    J., Lucas, T

    Weiss, D. J., Lucas, T. C. D., Nguyen, M., Nandi, A. K., Bisanzio, D., Battle, K. E., Cameron, E., Twohig, K. A., Pfeffer, D. A., Rozier, J. A., Gibson, H. S., Rao, P. C., Casey, D., Bertozzi-Villa, A., Collins, E. L., Dalrymple, U., Gray, N., Harris, J. R., Howes, R. E., Kang...

    2000

  29. [37]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.